add max_workers, textgrids
Browse files- libritts-aligned.py +110 -59
libritts-aligned.py
CHANGED
@@ -12,6 +12,7 @@ from alignments.datasets.libritts import LibrittsDataset
|
|
12 |
from tqdm.contrib.concurrent import process_map
|
13 |
from tqdm.auto import tqdm
|
14 |
from multiprocessing import cpu_count
|
|
|
15 |
from phones.convert import Converter
|
16 |
import torchaudio
|
17 |
import torchaudio.transforms as AT
|
@@ -22,7 +23,12 @@ _PHONESET = "arpabet"
|
|
22 |
|
23 |
_VERBOSE = os.environ.get("LIBRITTS_VERBOSE", True)
|
24 |
_MAX_WORKERS = os.environ.get("LIBRITTS_MAX_WORKERS", cpu_count())
|
|
|
25 |
_PATH = os.environ.get("LIBRITTS_PATH", os.environ.get("HF_DATASETS_CACHE", None))
|
|
|
|
|
|
|
|
|
26 |
if _PATH is not None and not os.path.exists(_PATH):
|
27 |
os.makedirs(_PATH)
|
28 |
|
@@ -56,6 +62,7 @@ _URLS = {
|
|
56 |
"train-clean-360": _URL + "train-clean-360.tar.gz",
|
57 |
"train-other-500": _URL + "train-other-500.tar.gz",
|
58 |
}
|
|
|
59 |
|
60 |
|
61 |
class LibriTTSAlignConfig(datasets.BuilderConfig):
|
@@ -72,11 +79,16 @@ class LibriTTSAlignConfig(datasets.BuilderConfig):
|
|
72 |
self.sampling_rate = sampling_rate
|
73 |
self.hop_length = hop_length
|
74 |
self.win_length = win_length
|
75 |
-
|
76 |
if _PATH is None:
|
77 |
-
raise ValueError(
|
|
|
|
|
78 |
elif _PATH == os.environ.get("HF_DATASETS_CACHE", None):
|
79 |
-
logger.warning(
|
|
|
|
|
|
|
80 |
|
81 |
class LibriTTSAlign(datasets.GeneratorBasedBuilder):
|
82 |
"""LibriTTSAlign dataset."""
|
@@ -99,7 +111,7 @@ class LibriTTSAlign(datasets.GeneratorBasedBuilder):
|
|
99 |
"phones": datasets.Sequence(datasets.Value("string")),
|
100 |
"phone_durations": datasets.Sequence(datasets.Value("int32")),
|
101 |
# audio feature
|
102 |
-
"audio": datasets.Value("string")
|
103 |
}
|
104 |
|
105 |
return datasets.DatasetInfo(
|
@@ -117,64 +129,98 @@ class LibriTTSAlign(datasets.GeneratorBasedBuilder):
|
|
117 |
ds_dict[name] = self._create_alignments_ds(name, url)
|
118 |
splits = [
|
119 |
datasets.SplitGenerator(
|
120 |
-
name=key.replace("-", "."),
|
121 |
-
|
122 |
-
)
|
123 |
for key, value in ds_dict.items()
|
124 |
]
|
125 |
# dataframe with all data
|
126 |
-
data_train =
|
127 |
-
|
128 |
-
|
129 |
-
|
130 |
-
|
131 |
-
|
132 |
-
|
133 |
-
|
134 |
-
"
|
135 |
-
|
136 |
-
|
137 |
-
|
138 |
-
|
139 |
-
|
140 |
-
|
141 |
-
|
142 |
-
|
143 |
-
|
144 |
-
|
145 |
-
|
146 |
-
|
147 |
-
|
148 |
-
|
149 |
-
|
150 |
-
|
151 |
-
|
152 |
-
|
153 |
-
|
154 |
-
|
155 |
-
|
156 |
-
|
157 |
-
|
158 |
-
|
159 |
-
|
160 |
-
|
161 |
-
|
162 |
-
|
163 |
-
)
|
164 |
-
|
165 |
-
|
166 |
-
|
167 |
-
"
|
168 |
-
|
169 |
-
|
170 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
171 |
self.alignments_ds = None
|
172 |
self.data = None
|
173 |
return splits
|
174 |
|
175 |
def _create_alignments_ds(self, name, url):
|
176 |
self.empty_textgrids = 0
|
177 |
-
ds_hash = hashlib.md5(
|
|
|
|
|
178 |
pkl_path = os.path.join(_PATH, f"{ds_hash}.pkl")
|
179 |
if os.path.exists(pkl_path):
|
180 |
ds = pickle.load(open(pkl_path, "rb"))
|
@@ -190,9 +236,11 @@ class LibriTTSAlign(datasets.GeneratorBasedBuilder):
|
|
190 |
target_directory=tgt_dir,
|
191 |
source_directory=src_dir,
|
192 |
source_url=url,
|
|
|
193 |
verbose=_VERBOSE,
|
194 |
tmp_directory=os.path.join(_PATH, f"{name}-tmp"),
|
195 |
chunk_size=1000,
|
|
|
196 |
)
|
197 |
pickle.dump(ds, open(pkl_path, "wb"))
|
198 |
return ds, ds_hash
|
@@ -209,7 +257,9 @@ class LibriTTSAlign(datasets.GeneratorBasedBuilder):
|
|
209 |
del data
|
210 |
for i, ds in enumerate(ds):
|
211 |
if os.path.exists(os.path.join(_PATH, f"{hashes[i]}-entries.pkl")):
|
212 |
-
add_entries = pickle.load(
|
|
|
|
|
213 |
else:
|
214 |
add_entries = [
|
215 |
entry
|
@@ -223,7 +273,10 @@ class LibriTTSAlign(datasets.GeneratorBasedBuilder):
|
|
223 |
)
|
224 |
if entry is not None
|
225 |
]
|
226 |
-
pickle.dump(
|
|
|
|
|
|
|
227 |
entries += add_entries
|
228 |
if self.empty_textgrids > 0:
|
229 |
logger.warning(f"Found {self.empty_textgrids} empty textgrids")
|
@@ -259,9 +312,7 @@ class LibriTTSAlign(datasets.GeneratorBasedBuilder):
|
|
259 |
if "[" not in phone:
|
260 |
o_phone = phone
|
261 |
if o_phone not in self.phone_cache:
|
262 |
-
phone = self.phone_converter(
|
263 |
-
phone, _PHONESET, lang=None
|
264 |
-
)[0]
|
265 |
self.phone_cache[o_phone] = phone
|
266 |
phone = self.phone_cache[o_phone]
|
267 |
phones.append(phone)
|
@@ -304,4 +355,4 @@ class LibriTTSAlign(datasets.GeneratorBasedBuilder):
|
|
304 |
"audio": str(row["audio"]),
|
305 |
}
|
306 |
yield j, result
|
307 |
-
j += 1
|
|
|
12 |
from tqdm.contrib.concurrent import process_map
|
13 |
from tqdm.auto import tqdm
|
14 |
from multiprocessing import cpu_count
|
15 |
+
import multiprocessing as mp
|
16 |
from phones.convert import Converter
|
17 |
import torchaudio
|
18 |
import torchaudio.transforms as AT
|
|
|
23 |
|
24 |
_VERBOSE = os.environ.get("LIBRITTS_VERBOSE", True)
|
25 |
_MAX_WORKERS = os.environ.get("LIBRITTS_MAX_WORKERS", cpu_count())
|
26 |
+
_MAX_WORKERS = int(_MAX_WORKERS)
|
27 |
_PATH = os.environ.get("LIBRITTS_PATH", os.environ.get("HF_DATASETS_CACHE", None))
|
28 |
+
_DOWNLOAD_SPLITS = os.environ.get(
|
29 |
+
"LIBRITTS_DOWNLOAD_SPLITS",
|
30 |
+
"train-clean-100,train-clean-360,train-other-500,dev-clean,dev-other,test-clean,test-other",
|
31 |
+
).split(",")
|
32 |
if _PATH is not None and not os.path.exists(_PATH):
|
33 |
os.makedirs(_PATH)
|
34 |
|
|
|
62 |
"train-clean-360": _URL + "train-clean-360.tar.gz",
|
63 |
"train-other-500": _URL + "train-other-500.tar.gz",
|
64 |
}
|
65 |
+
_URLS = {k: v for k, v in _URLS.items() if k in _DOWNLOAD_SPLITS}
|
66 |
|
67 |
|
68 |
class LibriTTSAlignConfig(datasets.BuilderConfig):
|
|
|
79 |
self.sampling_rate = sampling_rate
|
80 |
self.hop_length = hop_length
|
81 |
self.win_length = win_length
|
82 |
+
|
83 |
if _PATH is None:
|
84 |
+
raise ValueError(
|
85 |
+
"Please set the environment variable LIBRITTS_PATH to point to the LibriTTS dataset directory."
|
86 |
+
)
|
87 |
elif _PATH == os.environ.get("HF_DATASETS_CACHE", None):
|
88 |
+
logger.warning(
|
89 |
+
"Please set the environment variable LIBRITTS_PATH to point to the LibriTTS dataset directory. Using HF_DATASETS_CACHE as a fallback."
|
90 |
+
)
|
91 |
+
|
92 |
|
93 |
class LibriTTSAlign(datasets.GeneratorBasedBuilder):
|
94 |
"""LibriTTSAlign dataset."""
|
|
|
111 |
"phones": datasets.Sequence(datasets.Value("string")),
|
112 |
"phone_durations": datasets.Sequence(datasets.Value("int32")),
|
113 |
# audio feature
|
114 |
+
"audio": datasets.Value("string"),
|
115 |
}
|
116 |
|
117 |
return datasets.DatasetInfo(
|
|
|
129 |
ds_dict[name] = self._create_alignments_ds(name, url)
|
130 |
splits = [
|
131 |
datasets.SplitGenerator(
|
132 |
+
name=key.replace("-", "."), gen_kwargs={"ds": self._create_data(value)}
|
133 |
+
)
|
|
|
134 |
for key, value in ds_dict.items()
|
135 |
]
|
136 |
# dataframe with all data
|
137 |
+
data_train, data_dev, data_test, data_all = None, None, None, None
|
138 |
+
if (
|
139 |
+
"train-clean-100" in _URLS
|
140 |
+
and "train-clean-360" in _URLS
|
141 |
+
and "train-other-500" in _URLS
|
142 |
+
):
|
143 |
+
data_train = self._create_data(
|
144 |
+
[
|
145 |
+
ds_dict["train-clean-100"],
|
146 |
+
ds_dict["train-clean-360"],
|
147 |
+
ds_dict["train-other-500"],
|
148 |
+
]
|
149 |
+
)
|
150 |
+
if "dev-clean" in _URLS and "dev-other" in _URLS:
|
151 |
+
data_dev = self._create_data([ds_dict["dev-clean"], ds_dict["dev-other"]])
|
152 |
+
if "test-clean" in _URLS and "test-other" in _URLS:
|
153 |
+
data_test = self._create_data(
|
154 |
+
[ds_dict["test-clean"], ds_dict["test-other"]]
|
155 |
+
)
|
156 |
+
if (
|
157 |
+
"train-clean-100" in _URLS
|
158 |
+
and "train-clean-360" in _URLS
|
159 |
+
and "train-other-500" in _URLS
|
160 |
+
and "dev-clean" in _URLS
|
161 |
+
and "dev-other" in _URLS
|
162 |
+
and "test-clean" in _URLS
|
163 |
+
and "test-other" in _URLS
|
164 |
+
):
|
165 |
+
data_all = pd.concat([data_train, data_dev, data_test])
|
166 |
+
if data_all is not None:
|
167 |
+
splits.append(
|
168 |
+
datasets.SplitGenerator(
|
169 |
+
name="train.all",
|
170 |
+
gen_kwargs={
|
171 |
+
"ds": data_all,
|
172 |
+
},
|
173 |
+
)
|
174 |
+
)
|
175 |
+
if data_dev is not None:
|
176 |
+
splits.append(
|
177 |
+
datasets.SplitGenerator(
|
178 |
+
name="dev.all",
|
179 |
+
gen_kwargs={
|
180 |
+
"ds": data_dev,
|
181 |
+
},
|
182 |
+
)
|
183 |
+
)
|
184 |
+
if data_test is not None:
|
185 |
+
splits.append(
|
186 |
+
datasets.SplitGenerator(
|
187 |
+
name="test.all",
|
188 |
+
gen_kwargs={
|
189 |
+
"ds": data_test,
|
190 |
+
},
|
191 |
+
)
|
192 |
+
)
|
193 |
+
if data_dev is not None and data_all is not None:
|
194 |
+
# move last row for each speaker from data_all to dev dataframe
|
195 |
+
data_dev = data_all.copy()
|
196 |
+
data_dev = data_dev.sort_values(by=["speaker", "audio"])
|
197 |
+
data_dev = data_dev.groupby("speaker").tail(1)
|
198 |
+
data_dev = data_dev.reset_index()
|
199 |
+
# remove last row for each speaker from data_all
|
200 |
+
data_all = data_all[~data_all["audio"].isin(data_dev["audio"])]
|
201 |
+
splits += [
|
202 |
+
datasets.SplitGenerator(
|
203 |
+
name="train",
|
204 |
+
gen_kwargs={
|
205 |
+
"ds": data_all,
|
206 |
+
},
|
207 |
+
),
|
208 |
+
datasets.SplitGenerator(
|
209 |
+
name="dev",
|
210 |
+
gen_kwargs={
|
211 |
+
"ds": data_dev,
|
212 |
+
},
|
213 |
+
),
|
214 |
+
]
|
215 |
self.alignments_ds = None
|
216 |
self.data = None
|
217 |
return splits
|
218 |
|
219 |
def _create_alignments_ds(self, name, url):
|
220 |
self.empty_textgrids = 0
|
221 |
+
ds_hash = hashlib.md5(
|
222 |
+
os.path.join(_PATH, f"{name}-alignments").encode()
|
223 |
+
).hexdigest()
|
224 |
pkl_path = os.path.join(_PATH, f"{ds_hash}.pkl")
|
225 |
if os.path.exists(pkl_path):
|
226 |
ds = pickle.load(open(pkl_path, "rb"))
|
|
|
236 |
target_directory=tgt_dir,
|
237 |
source_directory=src_dir,
|
238 |
source_url=url,
|
239 |
+
textgrid_url=f"https://huggingface.co/datasets/cdminix/libritts-aligned/resolve/main/data/{name.replace('-', '_')}.tar.gz",
|
240 |
verbose=_VERBOSE,
|
241 |
tmp_directory=os.path.join(_PATH, f"{name}-tmp"),
|
242 |
chunk_size=1000,
|
243 |
+
n_workers=_MAX_WORKERS,
|
244 |
)
|
245 |
pickle.dump(ds, open(pkl_path, "wb"))
|
246 |
return ds, ds_hash
|
|
|
257 |
del data
|
258 |
for i, ds in enumerate(ds):
|
259 |
if os.path.exists(os.path.join(_PATH, f"{hashes[i]}-entries.pkl")):
|
260 |
+
add_entries = pickle.load(
|
261 |
+
open(os.path.join(_PATH, f"{hashes[i]}-entries.pkl"), "rb")
|
262 |
+
)
|
263 |
else:
|
264 |
add_entries = [
|
265 |
entry
|
|
|
273 |
)
|
274 |
if entry is not None
|
275 |
]
|
276 |
+
pickle.dump(
|
277 |
+
add_entries,
|
278 |
+
open(os.path.join(_PATH, f"{hashes[i]}-entries.pkl"), "wb"),
|
279 |
+
)
|
280 |
entries += add_entries
|
281 |
if self.empty_textgrids > 0:
|
282 |
logger.warning(f"Found {self.empty_textgrids} empty textgrids")
|
|
|
312 |
if "[" not in phone:
|
313 |
o_phone = phone
|
314 |
if o_phone not in self.phone_cache:
|
315 |
+
phone = self.phone_converter(phone, _PHONESET, lang=None)[0]
|
|
|
|
|
316 |
self.phone_cache[o_phone] = phone
|
317 |
phone = self.phone_cache[o_phone]
|
318 |
phones.append(phone)
|
|
|
355 |
"audio": str(row["audio"]),
|
356 |
}
|
357 |
yield j, result
|
358 |
+
j += 1
|